feat: context parallelism - #85

Open

jaisw7 wants to merge 44 commits into

Commits on Jan 2, 2026

feat: context parallelism
jaisw7
committed
d20c724
fix: CP integration test uses monkeypatch for env-independent defaults testing
jaisw7
committed
694b1a2
fix: add CP environment verification step in workflow
jaisw7
committed
f463b6f
fix: explicitly set DNET_CP_ENABLED=true in .env after make init
jaisw7
committed
2f925a4
feat: implement actual gRPC ring communication, address Copilot review feedback
jaisw7
committed
cfd46e6
feat: replace mock ring communicator with real gRPC implementation
jaisw7
committed
1f317f3
fix: update test_cp_ring_comm.py to use real gRPC instead of deleted mocks
jaisw7
committed
b9152e3
feat: add ContextParallelSettings to .env.example generation
jaisw7
committed
26c0641
fix: use macOS sed syntax (sed -i '') in CI workflow
jaisw7
committed
0e301b1
feat: implement strategy selection based on DNET_CP_ENABLED config
jaisw7
committed
a95ac19
feat: wire up CPAdapter in shard.py based on DNET_CP_ENABLED config
jaisw7
committed
63ec20c
feat: complete implementation plan compliance
jaisw7
committed
c7e95b6

Commits on Jan 3, 2026

fix: convert cross-proto imports to relative imports in pb2 files
jaisw7
committed
99f7e78
fix: filter out manager nodes from shards passed to topology solver
jaisw7
committed
1dac040
fix: robustly filter shards in CPTopologySolver
jaisw7
committed
d01ac5d
fix: implement CPAdapter execution loop to prevent deadlock
jaisw7
committed
013669a
refactor: Dynamic CP rank assignment and validation
jaisw7
committed
7ccbba5
fix(cp): register CPRingService on existing gRPC server, add model config
jaisw7
committed
0cf14a4
feat(cp): add CP launch script for multi-shard model loading
jaisw7
committed
1dff741
fix(cp): use DnetDeviceProperties from dnet_p2p for typed access
jaisw7
committed
fad866e
feat(api): add /v1/settings endpoint and typed CP utilities
jaisw7
committed
d3433e5
fix(scripts): add sys.path hack to resolve local imports
jaisw7
committed
53eeeb3
fix(api): remove incorrect validator for boolean logprobs field
jaisw7
committed
c33aa61
test(inference): remove obsolete logprobs validation test
jaisw7
committed
448a7b9
fix(scripts): improve error reporting in stress test
jaisw7
committed
0d81794
feat(cp): plumbing for max_position_embeddings override
jaisw7
committed
72458d4
fix(inference): increase token timeout to 1h for long context tests
jaisw7
committed
762a4b2
test(shard): fix mock object to include max_position_embeddings
jaisw7
committed
81f30b9
fix(test): increase client timeout to 3600s
jaisw7
committed
2d68dac
fix(shard): slice logits to last token to avoid OOM on long context
jaisw7
committed
7f80c2d
fix(cp): ensure only last CP rank samples to prevent race/OOM
jaisw7
committed
5bba4e5
fix(test): update mocks with CP rank fields to fix regression
jaisw7
committed
5f6fd78
fix(shard): revert CP rank check to resolve hang on small contexts
jaisw7
committed
1a6bf75
feat(cp): implement API multi-rank broadcast and restore CP rank check
jaisw7
committed
e0f10b1
fix(cp): send decode tokens only to last rank to avoid empty chunks
jaisw7
committed
4ca9758
fix(tests): fix 3 failing tests from CP changes
jaisw7
committed
7c87c7b
fix: remove dnet-tui submodule and add to gitignore
jaisw7
committed
2fc3b25
fix(cp): broadcast full tokens to all ranks for ring attention
jaisw7
committed
1a3389c
fix(needle-in-haystack): disable thinking for tests
jaisw7
committed
c6feb9b
fix(needle-in-haystack): include instructions to use non-thinking models for best results
jaisw7
committed
dc63fab
feat(cp): implement true ring attention with 1/N memory scaling
jaisw7
committed
23bf72f
fix(cp): prevent recursion error in wrapper and injection
jaisw7
committed
4fe4ab6

Commits on Jan 5, 2026

fix: numerical stability in distributed kv-cache
jaisw7
committed
35980fd

Commits on Jan 6, 2026

cleanup debug statements
jaisw7
committed
477fbee